VTM-Nav: Cross-Episode Object-Goal Navigation with Visual-Topological Memory

A new framework called VTM-Nav has been introduced by researchers for object-goal navigation across episodes without requiring training. This system employs a Hierarchical Visual-Topological Memory (VTM) that retains scene information throughout different episodes without the need for retraining. VTM organizes visual memories based on basic room layouts, differentiates between in-room and remote-visible data, and keeps cues for successful approaches. When a new task arises, VTM-Nav re-localizes the agent within the established scene structure, retrieves relevant records from likely rooms, and uses memory guidance to inform current candidates. This study tackles the issue faced by training-free ObjectNav agents, which often lose previously acquired scene knowledge after each task. VTM-Nav functions with fixed navigation components and model parameters, leveraging cross-episode experiences without relying on oracle data.